Matching and Recognition in 3D. Based on slides by Tom Funkhouser and Misha Kazhdan

Transcription

1 Matching and Recognition in 3D Based on slides by Tom Funkhouser and Misha Kazhdan

2 From 2D to 3D: Some Things Easier No occlusion (but sometimes missing data instead) Segmenting objects often simpler

3 From 2D to 3D: Many Things Harder Rigid transform has 6 degrees of freedom vs. 3 Brute-force algorithms much less practical Rotations do not commute Difficult to parameterize, search over No natural parameterization for surfaces in 3D Hard to do FFT, convolution, PCA Exception: range images

4 Shape Matching Challenge Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating 3D Query Shape Descriptor Best Matches 3D Database

5 Shape Matching Challenge Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating 3D Query Shape Descriptor Best Matches 3D Database

6 Shape Matching Challenge Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating 3D Query Shape Descriptor Best Matches 3D Database

7 Shape Matching Challenge Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating 3D Query Shape Descriptor Best Matches 3D Database

8 Shape Matching Challenge Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating 3D Query Shape Descriptor Best Matches 3D Database

9 Shape Matching Challenge Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating Invariant to transformations Invariant to deformations Insensitive to noise Insensitive to topology Robust to degeneracies Different Transformations (translation, scale, rotation, mirror)

10 Shape Matching Challenge Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating Invariant to transformations Invariant to deformations Insensitive to noise Insensitive to topology Robust to degeneracies Different Articulated Poses

11 Shape Matching Challenge Image courtesy of Ramamoorthi et al. Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating Invariant to transformations Invariant to deformations Insensitive to noise Insensitive to topology Robust to degeneracies Scanned Surface

12 Shape Matching Challenge Images courtesy of Viewpoint & Stanford Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating Invariant to transformations Invariant to deformations Insensitive to noise Insensitive to topology Robust to degeneracies Different Genus Different Tessellations

13 Shape Matching Challenge Images courtesy of Utah & De Espona Need shape descriptor & matching method that is: Concise to store Quick to compute Efficient to match Discriminating Invariant to transformations Invariant to deformations Insensitive to noise Insensitive to topology Robust to degeneracies No Bottom!

14 Taxonomy of 3D Matching Methods Images courtesy of Amenta & Osada Structural representations Skeletons Part-based methods Feature-based methods Statistical representations Attribute feature vectors Volumetric methods Surface-based methods View-based methods

15 Features on Surfaces Can construct edge and corner detectors Analogue of 1 st derivative: surface normal Analogue of 2 nd derivative: curvature Curvature at each point in each direction Minimum and maximum: principal curvatures Can threshold or do nonmaximum suppression

16 Taxonomy of 3D Matching Methods Image courtesy of Mao Chen Structural representations Skeletons Part-based methods Tables Feature-based methods Statistical representations Feature 1 Desks File cabinets Attribute feature vectors Volumetric methods Feature 2 Surface-based methods View-based methods

17 Example Shape distributions Shape representation: probability distributions Distance measure: difference between distributions Evaluation method: classification performance

18 Shape Distributions Key idea: map 3D surfaces to common parameterization by randomly sampling shape function 3D Models Randomly sample shape function Probability Probability Distance Distance D2 Shape Distributions Similarity Measure

19 Which Shape Function? Implementation: simple shape functions based on angles, distances, areas, and volumes θ θ A3 (angle) D1 (distance) D2 (distance) D3 (area) D4 (volume) [Ankerst 99]

20 D2 Shape Distribution Properties Concise to store? Quick to compute? Invariant to transforms? Efficient to match? Insensitive to noise? Insensitive to topology? Robust to degeneracies? Invariant to deformations? Discriminating?

21 D2 Shape Distribution Properties Concise to store? Quick to compute? Invariant to transforms? Efficient to match? Insensitive to noise? Insensitive to topology? Robust to degeneracies? Invariant to deformations? Discriminating? Probability Distance Skateboard 512 bytes (64 values) 0.5 seconds (10 6 samples)

22 D2 Shape Distribution Properties Concise to store Translation Quick to compute Rotation Invariant to transforms? { Mirror Efficient to match? Scale (w/ normalization) Insensitive to noise? Insensitive to topology? Robust to degeneracies? Invariant to deformations? Discriminating? Skateboard Normalized Means Porsche

23 D2 Shape Distribution Properties Concise to store Quick to compute Invariant to transforms Efficient to match? Insensitive to noise? Insensitive to topology? Robust to degeneracies? Invariant to deformations? Discriminating? Probability Porsche Distance Skateboard

24 D2 Shape Distribution Properties Concise to store Quick to compute Invariant to transforms Efficient to match Insensitive to noise? Insensitive to topology? Robust to degeneracies? Invariant to deformations? Discriminating? 1% Noise

25 D2 Shape Distribution Properties Concise to store Quick to compute Invariant to transforms Efficient to match Insensitive to noise Insensitive to topology Robust to degeneracies Invariant to deformations? Discriminating? Ellipsoids with Different Eccentricities

26 D2 Shape Distribution Properties Concise to store Quick to compute Invariant to transforms Efficient to match Insensitive to noise Insensitive to topology Robust to degeneracies Invariant to deformations Discriminating? Line Segment Cylinder Sphere Circle Cube Two Spheres

27 D2 Shape Distribution Results Question How discriminating are D2 shape distributions? Test database 133 polygonal models 25 classes 4 Mugs 6 Cars 3 Boats

28 D2 Shape Distribution Results D2 distributions are different across classes D2 shape distributions for 15 classes of objects

29 D2 Shape Distribution Results D2 distributions reveal gross shape of object Line Segment D2 shape distributions for 15 classes of objects

30 D2 Shape Distribution Results D2 distributions reveal gross shape of object Circle D2 shape distributions for 15 classes of objects

31 D2 Shape Distribution Results D2 distributions reveal gross shape of object Cylinder D2 shape distributions for 15 classes of objects

32 D2 Shape Distribution Results D2 distributions reveal gross shape of object Sphere D2 shape distributions for 15 classes of objects

33 D2 Shape Distribution Results D2 distributions reveal gross shape of object Two Spheres D2 shape distributions for 15 classes of objects

34 D2 Shape Distribution Results But are D2 distributions discriminating? D2 shape distributions for 15 classes of objects

35 D2 Shape Distribution Results Probability Distance D2 distributions for 5 tanks (gray) and 6 cars (black)

36 Evaluation Methods For each model (the query): Compute match score for all models Rank matches from best to worst Measure how often models in same class as query appear near top of ranked list Query Ranked Matches

37 Evaluation Methods Precision-recall curves Precision = retrieved_in_class / total_retrieved Recall = retrieved_in_class / total_in_class 100% 80% Precision 60% 40% 20% % 40% 60% 80% Recall 100%

38 Evaluation Methods Precision-recall curve example Precision = 0 / 0 Recall = 0 / Precision 100% 80% 60% 40% 20% Query Ranked Matches % 40% 60% 80% Recall 100%

39 Evaluation Methods Precision-recall curve example Precision = 1 / 1 Recall = 1 / Precision 100% 80% 60% 40% 20% Query Ranked Matches % 40% 60% 80% Recall 100%

40 Evaluation Methods Precision-recall curve example Precision = 2 / 3 Recall = 2 / Precision 100% 80% 60% 40% 20% Query Ranked Matches % 40% 60% 80% Recall 100%

41 Evaluation Methods Precision-recall curve example Precision = 3 / 5 Recall = 3 / Precision 100% 80% 60% 40% 20% Query Ranked Matches % 40% 60% 80% Recall 100%

42 Evaluation Methods Precision-recall curve example Precision = 4 / 7 Recall = 4 / Precision 100% 80% 60% 40% 20% Query Ranked Matches % 40% 60% 80% Recall 100%

43 Evaluation Methods Precision-recall curve example Precision = 5 / 9 Recall = 5 / Precision 100% 80% 60% 40% 20% Query Ranked Matches % 40% 60% 80% Recall 100%

44 Evaluation Methods 100% Precision 50% D2 Random 0% 0% 50% 100% Recall

45 Extended Gaussian Image [Horn, 1984] Represent a model by a spherical function by binning surface normals Model Angular Bins EGI

46 Extended Gaussian Image [Horn, 1984] Properties: Invertible for convex shapes Can be defined for most models 2D array of information Point Clouds Polygon Soups Closed Meshes Genus-0 Meshes Shape Spectrum

47 Extended Gaussian Image [Horn, 1984] Properties: Invertible for convex shapes Can be defined for most models 2D array of information Limitations: In general, shapes are not convex Normals are sensitive to noise 3D Model EGI

48 Extended Gaussian Image [Horn, 1984] Properties: Invertible for convex shapes Can be defined for most models 2D array of information Limitations: In general, shapes are not convex Normals are sensitive to noise Initial Model Noisy Model

49 Retrieval Results Princeton Shape Benchmark 51 potted plants 33 faces 15 desk chairs 22 dining chairs 100 humans 28 biplanes 14 flying birds 11 ships

50 Retrieval Results 100% Precision 50% Extended Gaussian Image D2 Random 0% 0% 50% 100% Recall

51 Shape Histograms [Ankerst et al., 1999] Shape descriptor stores a histogram of how much surface resides at different bins in space Model Shape Histogram (Sectors + Shells)

52 Boundary Voxel Representation Represent a model as the (anti-aliased) rasterization of its surface into a regular grid: A voxel has value 1 (or area of intersection) if it intersects the boundary A voxel has value 0 if it doesn t intersect Model Voxel Grid

53 Boundary Voxel Representation Properties: Can be defined for any model Invertible 3D array of information Point Clouds Polygon Soups Closed Meshes Genus-0 Meshes Shape Spectrum

54 Boundary Voxel Representation Properties: Can be defined for any model Invertible 3D array of information Limitations: Difficult to match If the resolution is too high: most voxels miss If the resolution is too low: representation is too coarse Intersection

55 Retrieval Results Precision 100% 50% Shape Histograms Extended Gaussian Image D2 Random 0% 0% 50% 100% Recall

56 Histogram Representations Challenge: If shape properties are mapped to nearby bins, they will not be compared Solutions: Match across adjacent bins: Earth Mover s Distance Low-pass filter: Convolution with a Gaussian

57 Earth Mover s distance [Rubner et al. 1998] Match by computing the minimal amount of work needed to transform one distribution into the other. Computing the distance: For 1D histograms can use the CDF to compare efficiently In general, need to solve the transportation problem which is inefficient for large numbers of bins

58 Convolving with a Gaussian The value at a point is obtained by summing Gaussians distributed over the surface of the model. Distributes the surface into adjacent bins Blurs the model, loses high frequency information Surface Gaussian Gaussian convolved surface

59 Gaussian EDT [Kazhdan et al., 2003] The value at a point is obtained by summing the Gaussian of the closest point on the model surface. Distributes the surface into adjacent bins Maintains high-frequency information max Surface Gaussian Gaussian EDT

60 Gaussian EDT Properties: Can be defined for any model Invertible 3D array of information Difference measures proximity between surfaces Point Clouds Polygon Soups Closed Meshes Genus-0 Meshes Shape Spectrum

61 Retrieval Results Precision 100% 50% Gaussian EDT Shape Histograms Extended Gaussian Image D2 Random 0% 0% 50% 100% Recall

62 Handling Transformations Key difficulty: locating objects under any rigid-body transformation Approaches: Exhaustive search: try all possibilities Invariance: use descriptors that do not change under transformations Normalization: align objects to canonical coordinate frame

63 Exhaustive Search Search for the best aligning transformation: Compare at all alignments Match at the alignment for which models are closest Exhaustive search for optimal rotation

64 Exhaustive Search Search for the best aligning transformation: Compare at all alignments Match at the alignment for which models are closest

65 Exhaustive Search Search for the best aligning transformation: Use signal processing for efficient correlation Represent model at many different transformations Search for the best aligning transformation: Gives the correct answer Is hard to do efficiently

66 Invariance Represent a model with information that is independent of the transformation Power spectrum representation Fourier Transform for translation and 2D rotations Spherical Harmonic Transform for 3D rotations Frequency Circular Power Spectrum Frequency Spherical Power Spectrum

67 Circular Power Spectrum Circular Function

68 Circular Power Spectrum + = Circular Function Cosine/Sine Decomposition

69 Circular Power Spectrum + = Circular Function = Constant Frequency Decomposition

70 Circular Power Spectrum Circular Function + = = + Constant 1 st Order Frequency Decomposition

71 Circular Power Spectrum Circular Function + = = + + Constant 1 st Order 2 nd Order Frequency Decomposition

72 Circular Power Spectrum Circular Function + = = Constant 1 st Order 2 nd Order 3 rd Order Frequency Decomposition

73 Circular Power Spectrum Circular Function = Amplitudes invariant to rotation = Constant 1 st Order 2 nd Order 3 rd Order Frequency Decomposition

74 Spherical Power Spectrum Represent each spherical function as a sum of harmonic frequencies (orders) = Harmonic Decomposition

75 Spherical Power Spectrum Represent each spherical function as a sum of harmonic frequencies (orders) = Constant 1 st Order 2 nd Order 3 rd Order + + +

76 Spherical Power Spectrum Store how much (L 2 -norm) of the shape resides in each frequency to get a rotation invariant representation Constant 1 st Order 2 nd Order 3 rd Order = + + +

77 Invariant to transforms? Frequency subspaces are fixed by rotations:

78 Invariant to transforms? Frequency subspaces are fixed by rotations:

79 Invariant to transforms? Frequency subspaces are fixed by rotations:

80 Invariant to transforms? Frequency subspaces are fixed by rotations:

81 Power Spectrum Translation-invariance: Represent the model in a Cartesian coordinate system Compute the 3D Fourier transform Store the amplitudes of the frequency components f ( x, Cartesian Coordinates y, z) = l, m, n f l e, m, n i( lx+ my+ zn) y x z { f } l, m, n l, m, n Translation Invariant Representation

82 Power Spectrum Single axis rotation-invariance: Represent the model in a cylindrical coordinate system Compute the Fourier transform in the angular direction Store the amplitudes of the frequency components η r h f Cylindrical Coordinates ( r, h, θ ) = k f k { f } k ( r, h) k Rotation Invariant Representation ( r, h) e i( kθ )

83 Power Spectrum Full rotation-invariance: Represent the model in a spherical coordinate system Compute the spherical harmonic transform Store the amplitudes of the frequency components η ϖ Spherical Coordinates m f ( r, θ, φ) = f, ( r) Y ( θ, φ) l m l l m l r m l m f l r ( ) 2 Rotation Invariant Representation l

84 Power Spectrum Power spectrum representations Are invariant to transformations Give a lower bound for the best match Tend to discard too much information Translation invariant: n 3 data -> n 3 /2 data Single-axis rotation invariant: n 3 data -> n 3 /2 data Full rotation invariant: n 3 data -> n 2 data

85 Normalization Place a model into a canonical coordinate frame by normalizing for: translation scale rotation Translation Rotation Scale

86 Alignment of Point Sets [Horn et al., 1988] Given two point sets P={p 1,,p n } and Q={q 1,,q n }, what is the transformation T minimizing the sum of squared distances: d( P, Q) = n i= 1 p T ( q ) i i 2 p i q n q1 q i q2 p n p 1 p 2 Point set P Point set Q

87 Alignment of Point Sets [Horn et al., 1988] Translation Align the models so that their center of mass is at the origin. n n i= 1 p i = 0 and q i= 1 i = 0 p i q i p 2 p n q n q 1 q 2 p 1 Translated point set P Translated point set Q

88 Alignment of Point Sets [Horn et al., 1988] Scale Align the models so that their mean variance is 1. n i= 1 2 p i = 1 and q n i= 1 i 2 = 1 p i q i p n p 1 p 2 q n q 1 q 2 Scaled point set P Scaled point set Q

89 Alignment of Point Sets [Horn et al., 1988] Rotation Factor the cross covariance matrix: ( p p ) ( q ) t M = n 1 q n p i q i p n p 1 p 2 q n q 1 q 2 Rotationally aligned point sets P and Q

90 Normalization Place a model into a canonical coordinate frame: Translation: center of mass i= 1 Can be done on a per-model basis Scale: mean variance n i= 1 n p i = 0 and q 2 i= 1 Can be done on a per-model basis Rotation: factoring cross covariance matrix Need to know the correspondences between models n p i = 1 and q n i= 1 ( p p ) ( q ) t M = n 1 q n i i = 2 0 = 1

91 Rotation Challenge: We want to normalize for rotation on a per-model basis Solution: Align the model so that the principal axes align with the coordinate axes PCA Alignment

92 Rotation Challenge: We want to normalize for rotation on a per-model basis Solution: Align the model so that the principal axes align with the coordinate axes PCA Alignment Directions of the axes are ambiguous

93 Normalization (PCA) PCA defines a coordinate frame up to reflection in the coordinate axes. y Make descriptor invariant to the eight reflections Reflections fix the cosine term Reflections multiply the sine term by -1 f ( θ ) = a cos( kθ ) + b sin( kθ ) k k k x z { } k k b k a, Translation Invariant Representation

94 Problem with PCA-Based Alignment If singular values are close, axes unstable

95 Retrieval Results (Rotation) 100% 50% Gaussian EDT Exhaustive Search PCA + Flip Invariance Cylindrical Power Spectrum Spherical Power Spectrum PCA Size: Method Floats Exhaustive Search 8192 PCA + Flip Invariance 8192 PCA 8192 Cylindrical PS 4352 Spherical PS 512 0% 0% 50% 100% Recall Time: Method Secs. Exhaustive Search PCA + Flip Invariance.67 PCA.67 Cylindrical PS.32 Spherical PS.03